基于多分支注意力增强的细粒度图像分类

Fine-grained Image Classification Based on Multi-branch Attention-augmentation

针对细粒度图像类间差距小、类内差距大的问题,文中提出以弱监督学习的方式使用多分支注意力增强卷积网络,从而实现细粒度图像分类。文中采用Inception-V3网络提取图像的基础特征,从中获取多个局部响应区域并进行特征融合,在此基础上采用注意力机制对图像关键区域进行自约束的局部裁剪和局部擦除,避免仅提取目标单个部位的特征,促使网络更加关注目标物体不同部位的细节特征,同时也提升了目标区域的定位精度。此外,文中提出中心正则化损失函数来约束训练过程中获取的注意力区域,以进一步提升目标定位精度和扩大图像特征的类间差距。在3个公开数据集上进行了实验,结果表明,所提方法取得了比当前最优方法更好的结果。

In order to address the challenges of high intra-class variances and low inter-class variances in fine-grained image classification,a multi-branch attention-augmented convolution neural network is proposed to solve the problem.The pre-trained Inception-V3 network is used to extract basic feature.In order to solve the problem that features are extracted from one part of an object and encourage the network to pay more attention to the discriminative features of different parts,we apply self-constrained attention-wised cropping and self-constrained attention-wised erasing on the central parts of the original images.It also improves the detection accuracy of object locations.Meanwhile,a central regularization loss function is proposed to constrain attention-augmented training process to obtain better attention regions and expand the gap between different classes of images.Comprehensive experiments on three benchmark datasets show that our approach surpasses the state-of-art works.